The composition of the output layer architecture in a back-propagation (BP) neural network for remote sensing image classification

A Neural Network is treated as a data transformer when used for mapping purposes. The objective in this case, is to associate the elements in one set of data with the elements in a second set. According to this principle, three encoding methods, namely, single output layer, binary encoding, and ortho-encoding, have been designed for the output layer of a Neural Network based on five criteria, and put into experiments for Remote Sensing classification by means of a series of images coordinated with incremental noise level, from 1% to 10%. At last, the experiment results are assessed from different perspectives such as, accuracy, convergence, mixture detection, and confidence of classification, comparing to three encoding methods respectively.